Large cell carcinoma is a rare and aggressive lung cancer subtype. Large cell carcinoma (LCC) is a distinct subtype of non-small cell lung cancer characterized by undifferentiated, large polygonal cells that lack the defining features of adenocarcinoma or squamous cell carcinoma. This makes LCC a diagnosis often reached by exclusion. Despite representing a smaller proportion of lung cancers, LCC is associated with rapid progression, early metastasis, and poor prognosis, with smoking as the primary risk factor.
Histopathological ambiguity limits treatment options. LCC lacks well-defined molecular markers and targetable drivers, complicating diagnosis and limiting the effectiveness of targeted therapies. In the absence of a standard LCC-specific treatment protocol, therapeutic decisions are typically guided by broader NSCLC treatment strategies, meaning patients do not receive subtype-optimized care.
AJCC staging fails to capture LCC's unique biology. Traditional staging systems like the AJCC TNM framework depend on tumor-node-metastasis features that are well-suited for common subtypes like adenocarcinoma and squamous cell carcinoma. However, these models fail to capture the unique biological behavior of LCC, including its unpredictable progression and absence of targetable molecular drivers. Clinicians currently lack reliable tools to stratify LCC patients into risk groups or predict outcomes accurately.
No machine learning prognostic model existed for LCC at population scale. Despite broad advances in applying machine learning to cancer prognosis, no prior study had developed an ML-based prognostic model for LCC using large-scale population data. This gap underscores the need for a robust, population-based tool specifically tailored to this aggressive and understudied subtype. This study addresses that gap using the SEER database and multiple machine learning algorithms.
1,867 LCC patients from the SEER database. Of 3,676 patients initially screened from the Surveillance, Epidemiology, and End Results (SEER) program, 1,867 met predefined inclusion criteria with complete demographic, clinicopathological, treatment, and survival data. LCC was defined using SEER ICD-O-3 histology codes covering large cell carcinoma NOS, large cell neuroendocrine carcinoma, lymphoepithelioma-like carcinoma, basaloid carcinoma, and clear cell carcinoma. Overall survival, from diagnosis to death from any cause, was the primary endpoint.
Two-step feature selection: Lasso-Cox then Boruta. A rigorous two-step feature selection strategy was applied entirely within the training set to prevent data leakage. First, Lasso-Cox regression using the lambda.1se criterion performed dimensionality reduction from all available demographic, tumor, metastatic, and treatment variables. Second, the Boruta algorithm was applied to the Lasso-selected candidates to further refine the feature set, retaining only the most informative and stable predictors while minimizing redundancy.
Nine prognostic features identified. Boruta analysis identified nine key prognostic factors: chemotherapy, tumor size, T stage, age, M stage, surgery of the primary site, N stage, and metastasis to the liver and brain. Lasso-Cox regression had excluded race, marital status, tumor grade, tumor laterality, radiotherapy, and metastasis to bone and lung as independent predictors of overall survival.
SMOTE and 7:3 train-validation split. Patients were randomly divided into a training set of 70% and a validation set of 30%. To address class imbalance (72.9% mortality rate), Synthetic Minority Over-sampling Technique (SMOTE) was applied to balance the outcome labels in the training set. All hyperparameter tuning was conducted using 10-fold cross-validation within the training set, with the validation set held out entirely for independent performance evaluation.
Four ML algorithms compared against Cox regression and AJCC staging. Using the nine selected prognostic features, four machine learning models were trained: Random Survival Forest (RSF), XGBoost (XGB), Artificial Neural Network (ANN), and Decision Tree (DT). These were benchmarked against traditional Cox proportional hazards regression and the AJCC staging system. Each ML model underwent systematic hyperparameter tuning via parallelized grid search with a fixed random seed for reproducibility.
RSF hyperparameters tuned across key tree parameters. The RSF model was tuned for number of trees (200 to 800), number of variables per split (3 to 5), and minimum node size (15 to 21), using a grid search with resolution of 3. XGBoost tuning covered boosting iterations, tree depth, and learning rate over 60 evaluations. ANN tuning included dropout rate, weight decay, hidden layer count, nodes per layer, and learning rate. Decision tree optimization covered splitting significance levels and minimum leaf observations.
Comprehensive evaluation using four performance dimensions. Model performance was assessed across four complementary dimensions: discrimination via time-dependent AUC (for 3-year and 5-year OS); calibration via calibration plots comparing predicted to observed outcomes; clinical utility via decision curve analysis (DCA) measuring net benefit across threshold probabilities; and predictive accuracy via Brier scores quantifying squared differences between predicted and observed survival probabilities.
SHAP, partial dependence, and restricted cubic splines for interpretation. Model interpretability was addressed through SHAP beeswarm plots identifying feature importance and direction of effect, breakdown analysis for patient-level prediction explanation, partial dependence plots (PDPs) showing feature effects on survival over time, and Restricted Cubic Spline (RCS) analysis to detect non-linear relationships between continuous features and survival outcome.
RSF achieved the highest discriminatory performance. The Random Survival Forest model demonstrated superior discriminatory ability compared to all evaluated models. In the training dataset, RSF achieved time-dependent AUC values of 0.858 (95% CI 0.838 to 0.878) for 3-year overall survival prediction and 0.863 (95% CI 0.840 to 0.886) for 5-year overall survival prediction. This strong performance was maintained in the validation dataset with AUCs of 0.841 (95% CI 0.808 to 0.874) for 3-year and 0.808 (95% CI 0.764 to 0.845) for 5-year prediction.
RSF outperformed Cox regression and AJCC staging on all metrics. Traditional Cox regression achieved training AUCs of 0.814 and 0.810 for 3-year and 5-year OS, respectively, with validation AUCs of 0.828 and 0.805. The AJCC staging model achieved the lowest discrimination with training AUCs of 0.788 and 0.757. The ANN performed poorly in external validation (AUC 0.615 for 3-year, 0.629 for 5-year), showing substantial overfitting. XGB and DT were competitive with Cox regression but inferior to RSF across most metrics.
RSF achieved the lowest Brier scores. The RSF model achieved Brier scores of 0.150 and 0.137 for 3-year and 5-year predictions in the training set, and 0.158 and 0.162 in the validation set. These were the lowest scores among all models evaluated, indicating the most accurate survival probability estimates. Cox regression achieved Brier scores of 0.168 and 0.156 in training and 0.162 and 0.158 in validation, while ANN showed dramatically worse Brier scores of 0.564 and 0.635 in external validation.
RSF showed superior calibration and net clinical benefit. Calibration analysis confirmed that RSF's predicted survival probabilities most closely aligned with observed outcomes, outperforming all other models. Decision curve analysis demonstrated that RSF provided greater net clinical benefit than Cox regression, AJCC staging, and other ML models across a wide range of threshold probabilities, supporting its practical utility for individualized risk stratification in LCC patients.
Tumor size is the most influential prognostic feature. SHAP analysis of the RSF model identified tumor size as the feature exerting the greatest influence on mortality risk, with higher values strongly associated with elevated hazard. This was followed in order by liver metastasis, M stage, brain metastasis, T stage, chemotherapy status, surgery of the primary site, N stage, and age. Among categorical features, M1 status, T4 classification, and presence of liver metastases were strongly associated with elevated mortality risk via positive SHAP values.
Brain and liver metastases are the strongest categorical predictors of poor survival. Partial dependence profiles confirmed that brain and liver metastases were associated with the steepest survival probability declines over time. M1 patients experienced worse outcomes than M0 patients, and survival probabilities declined progressively as N stage advanced, with N3 patients experiencing the poorest nodal outcomes. Tumor T stage followed a similar gradient from T1 (most favorable) to T4 (poorest).
Surgery and chemotherapy improve survival; absence of treatment is a key risk factor. Treatment-related partial dependence plots showed that patients who underwent surgery at the primary tumor site demonstrated significantly improved survival compared to those who did not. Chemotherapy administration was also associated with a survival benefit, while absence of known chemotherapy treatment correlated with poorer prognosis. Patient-level breakdown analysis of an example patient showed that surgical intervention and early AJCC stage contributed favorably, while absence of chemotherapy was linked to worse predicted outcomes.
Non-linear relationship identified between tumor size and mortality. RCS analysis revealed a statistically significant non-linear relationship between tumor size and mortality risk (p = 0.023), with a threshold identified at 53.47 mm. Below this threshold, the relationship between tumor size and mortality risk was more gradual; above it, the association steepened substantially. In contrast, the non-linear relationship between age and mortality was not statistically significant (p greater than 0.05), suggesting that age exerts a more linear influence on LCC prognosis.
An interactive web application translates the RSF model into clinical practice. Building on the RSF model's superior performance, the study developed a publicly accessible web-based prognostic application hosted at the Shiny platform. The application allows clinicians to enter patient-level values for the nine identified prognostic features and receive individualized overall survival probability predictions at 3-year and 5-year time horizons, along with dynamic risk stratification visualizations.
Designed for clinicians without machine learning expertise. The platform provides an intuitive and automated framework for model setup and evaluation that does not require users to understand the underlying RSF algorithm. By translating complex survival model outputs into clinically interpretable risk profiles, the tool enables data-driven treatment planning and follow-up scheduling decisions that would otherwise require specialized computational expertise.
Addresses the absence of an LCC-specific prognostic standard. Because LCC lacks standard treatment protocols and is typically managed using general NSCLC guidelines, individualized prognostic assessment has been particularly difficult. The RSF-based web tool fills this gap by providing LCC-specific survival predictions that incorporate the disease's unique prognostic drivers, enabling clinicians to customize treatment sequencing and post-surgical surveillance for individual patients.
Clinical utility supported across diverse risk thresholds. Decision curve analysis confirmed that the RSF model provides net clinical benefit exceeding both the treat-all and treat-none strategies across a broad range of probability thresholds. This finding means the tool is clinically useful not only for patients at clearly high or low risk but also across the intermediate range where treatment decisions are most uncertain and where individualized prognostic guidance adds the greatest value.
AUC overlap with Cox regression warrants multi-metric evaluation. Although the RSF model showed slightly higher time-dependent AUC values than Cox regression, partial overlap in confidence intervals was observed, meaning the discrimination advantage was not definitively statistically significant when considered in isolation. The study addressed this by evaluating performance across multiple complementary metrics. RSF's consistent superiority on calibration, Brier score, and DCA across both training and validation datasets supports its overall advantage even when AUC differences alone are ambiguous.
SEER histologic classification introduces heterogeneity. The study used SEER ICD-O-3 coding, which groups multiple related subtypes under the LCC category for registry purposes. The contemporary WHO classification recognizes several of these entities, including large cell neuroendocrine carcinoma, as distinct pathological subtypes with different biology and treatment implications. This approach may introduce cohort heterogeneity and limit clinical interpretability for strictly defined LCC populations, though it enhances representativeness within real-world registry data.
SEER database lacks molecular and treatment detail. The SEER database does not include detailed molecular profiles such as PIK3A, CDKN2A, or TP53 mutation status, smoking history, comorbidities, or specific treatment regimens beyond receipt or non-receipt of surgery, chemotherapy, and radiotherapy. These omissions limit the model's ability to incorporate biological and treatment-specific nuances that are increasingly recognized as important in LCC prognosis. External validation using independent multicenter datasets with richer clinical data is essential.
Prior LCC prognostic research lacked ML-based models. Previous studies by Yao et al., Liu et al., Zheng et al., and Hu et al. identified prognostic factors in LCC subpopulations using traditional statistical methods such as Cox regression and Kaplan-Meier analysis. However, these studies did not develop comprehensive prognostic models tailored to LCC and focused on specific subpopulations such as elderly patients or resected cases. The current study represents the first ML-based prognostic model for LCC developed from a large-scale population dataset.
RSF outperforms all tested models for LCC overall survival prediction. Among six models evaluated including Cox regression, AJCC staging, XGBoost, ANN, and Decision Tree, the Random Survival Forest demonstrated the strongest discriminatory power, calibration, and clinical utility for predicting overall survival in large cell carcinoma. This establishes RSF as the preferred algorithmic approach for this rare and aggressive NSCLC subtype.
Nine clinical features drive LCC prognosis. The final model identified tumor size, liver metastasis, M stage, brain metastasis, T stage, chemotherapy, surgery of the primary site, N stage, and age as the nine key prognostic features. Tumor size was the single most influential predictor, with a statistically significant non-linear threshold at 53.47 mm, providing a potential clinical reference point for risk classification that goes beyond simple TNM staging categories.
A web-based tool brings LCC-specific prediction to clinical practice. By deploying the RSF model as a publicly accessible web application, this study delivers LCC-specific prognostic capability to clinicians who lack computational expertise. The tool enables individualized risk stratification, personalized treatment planning, and optimized follow-up scheduling for a patient population that has historically been managed without disease-specific prognostic guidance.
External validation is the critical next step. Although the study used rigorous internal validation methods including 10-fold cross-validation and a held-out validation set, external validation using independent multicenter datasets remains essential before routine clinical implementation. Future work should also incorporate molecular biomarkers, detailed treatment data, and comorbidity information to improve the model's biological comprehensiveness and translational applicability for LCC patients.